1. Field of the Invention
The present invention is generally related to the design of RISC type microprocessor architectures and, in particular, to RISC microprocessor architectures that are capable of executing multiple instructions concurrently.
2. Background
Recently, the design of microprocessor architectures have matured from the use of Complex Instruction Set Computer (CISC) to simpler Reduced Instruction Set Computer (RISC) Architectures. The CISC architectures are notable for the provision of substantial hardware to implement and support an instruction execution pipeline. The typical conventional pipeline structure includes, in fixed order, instruction fetch, instruction decode, data load, instruction execute and data store stages. A performance advantage is obtained by the concurrent execution of different portions of a set of instructions through the respective stages of the pipeline. The longer the pipeline, the greater the number of execution stages available and the greater number of instructions that can be concurrently executed.
Two general problems limit the effectiveness of CISC pipeline architectures. The first problem is that conditional branch instructions may not be adequately evaluated until a prior condition code setting instruction has substantially completed execution through the pipeline.
Thus, the subsequent execution of the conditional branch instruction is delayed, or stalled, resulting in several pipeline stages remaining inactive for multiple processor cycles. Typically, the condition codes are written to a condition code register, also referred to as a processor status register (PSR), only at completion of processing an instruction through the execution stage. Thus, the pipeline must be stalled with the conditional branch instruction in the decode stage for multiple processor cycles pending determination of the branch condition code. The stalling of the pipeline results in a substantial loss of through-put. Further, the average through-put of the computer will be substantially dependent on the mere frequency of conditional branch instructions occurring closely after the condition code setting instructions in the program instruction stream.
A second problem arises from the fact that instructions closely occurring in the program instruction stream will tend to reference the same registers of the processor register file. Data registers arc often used as the destination or source of data in the store and load stages of successive instructions. In general, an instruction that stores data to the register file must complete processing through at least the execution stage before the load stage processing of a subsequent instruction can be allowed to access the register file. Since the execution of many instructions require multiple processor cycles in the single execution stage to produce store data, the entire pipeline is typically stalled for the duration of an execution stage operation. Consequently, the execution through-put of the computer is substantially dependent on the internal order of the instruction stream being executed.
A third problem arises not so much from the execution of the instructions themselves, but the maintenance of the hardware supported instruction execution environment, or state-of-the-machine, of the microprocessor itself. Contemporary CISC microprocessor hardware sub-systems can detect the occurrence of trap conditions during the execution of instructions. Traps include hardware interrupts, software traps and exceptions. Each trap requires execution of a corresponding trap handling routines by the processor. On detection of the trap, the execution pipeline must be cleared to allow the immediate execution of the trap handling routine. Simultaneously, the state-of-the-machine must be established as of the precise point of occurrence of the trap; the precise point occurring at the conclusion of the first currently executing instruction for interrupts and traps and immediately prior to an instruction that fails due to a exception. Subsequently, the state-of-the-machine and, again depending on the nature of the trap the executing instruction itself must be restored at the completion of the handling routine. Consequently, with each trap or related event, a latency is introduced by the clearing of the pipeline at both the inception and conclusion of the handling routine and storage and return of the precise state-of-the-machine with corresponding reduction in the through-put of the processor.
These problems have been variously addressed in an effort to improve the potential through-put of CISC architectures. Assumptions can be made about the proper execution of conditional branch instructions, thereby allowing pipeline execution to tentatively proceed in advance of the final determination of the branch condition code. Assumptions can also be made as to whether a register will be modified, thereby allowing subsequent instructions to also be tentatively executed. Finally, substantial additional hardware can be provided to minimize the occurrence of exceptions that require execution of handling routines and thereby reduce the frequency of exceptions that interrupt the processing of the program instruction stream.
These solutions, while obviously introducing substantial additional hardware complexities, also introduce distinctive problems of their own. The continued execution of instructions in advance of a final resolution of either a branch condition or register file store access require that the state-of-the-machine be restorable to any of multiple points in the program instruction stream including the location of the conditional branch, each modification of a register file, and for any occurrence of an exception; potentially to a point prior to the fully completed execution of the last several instructions. Consequently, even more supporting hardware is required and, further, must be particularly designed not to significantly increase the cycle time of any pipeline stage.
RISC architectures have sought to avoid many of the foregoing problems by drastically simplifying the hardware implementation of the microprocessor architecture. In the extreme, each RISC instruction executes in only three pipelined program cycles including a load cycle, an execution cycle, and a store cycle. Through the use of load and store data bypassing, conventional RISC architectures can essentially execute a single instruction per cycle in the three stage pipeline.
Whenever possible, hardware support in RISC architectures is minimized in favor of software routines for performing the required functions. Consequently, the RISC architecture holds out the hope of substantial flexibility and high speed through the use of a simple load/store instruction set executed by an optimally matched pipeline. And in practice, RISC architectures have been found to benefit from the balance between a short, high-performance pipeline and the need to execute substantially greater numbers of instructions to implement all required functions.
The design of the RISC architecture generally avoids or minimizes the problems encountered by CISC architectures with regard to branches, register references and exceptions. The pipeline involved in a RISC architecture is short and optimized for speed. The shortness of the pipeline minimizes the consequences of a pipeline stall or clear as well as minimizing the problems in restoring the state-of-the-machine to an earlier execution point.
However, significant through-put performance gains over the generally realized present levels cannot be readily achieved by the conventional RISC architecture. Consequently, alternate, so-called superscalar architectures, have been variously proposed. These architectures generally attempt to execute multiple instructions concurrently and thereby proportionately increase the through-put of the processor. Unfortunately, such architectures are, again, subject to similar, if not the same conditional branch, register referencing, and exception handling problems as encountered by CISC architectures.